Graph attention network predicts drug-gene associations of matrix metalloproteinases 9-based host modulation in periodontitis

Abstract

Introduction:

Matrix metalloproteinases (MMPs) are essential endopeptidases involved in matrix degradation and remodeling, including periodontal tissues. They are classified into collagenases, gelatinases, stromelysin, matrilysin, and membrane types. MMPs, particularly MMP-2 and 9, contribute to gingival tissue breakdown in periodontitis. The study uses Graph Attention Network (GAT) to predict drug-gene associations for MMP-9 in host modulation, a crucial aspect of disease diagnosis, prognosis, targeted therapies, personalized medicine, and mechanistic studies. This approach can optimize treatment outcomes and minimize side effects, contributing to precision medicine.

Materials and Methods:

Data on drugs and genes associated with MMP-9 were retrieved using probes and drugs, and 1898 drug-gene interactions were studied. Data were cleaned for missing values, and graph data were prepared using nodes, gene names, and edges. Edge weights represented biochemical activity, while node features provided additional details for training a GAT. Cytoscape was used to create a network graph for drug-gene associations, while Cytohubba applied the maximum clique centrality algorithm to a drug-gene interaction network. A GAT model, consisting of three layers, was applied using Google Colab in a Python environment.

Results:

The network graph has 742 nodes, 1897 edges, and an average number of neighbors of 5.049. It has a characteristic path length of 3.303, with low local connectivity, and sparseness. The top-ten hubs with drug-gene associations with MMP-9 include quercetin, luteolin, econazole, zinc chloride, curcumin, MMP-9, MMP2, MMP1, MMP13, and MMP3. The model faces issues due to a dataset imbalance, with 80% of positive cases overfitting the majority class. Despite this, it learns useful features from the graph structure and shows stable training. The GAT model achieved an accuracy of 0.7955, indicating 80% correct classification, and an F1 score of 0.8861.

Conclusion:

This study explores the intricate relationship between drugs, genes, and MMP-9, using a GAT tool to identify potential drug targets. Addressing limitations can advance MMP-9 biology and develop new therapeutic strategies.

INTRODUCTION

Matrix metalloproteinases (MMPs) are a heterogeneous group of endopeptidases that play a crucial role in matrix degradation and tissue remodeling, including that of periodontal tissues.[1] By degrading the extracellular matrix components, these enzymes participate in the remodeling process that helps maintain the periodontal tissues.[2] Woessner discovered these enzymes in 1962 when he identified an enzyme capable of degrading collagen in the mammalian uterus.[3,4,5,6,7] Subsequently, several MMPs were identified, isolated, and systematically classified and numbered. The fundamental classes of these enzymes are grouped as collagenases, gelatinases, stromelysin, matrilysin, and the membrane-types.[8] Initially secreted as inactive precursors called zymogens, they are activated in response to functional demands. Apart from their crucial engagement in periodontal tissue homeostasis, specific MMPs can exhibit destructive behavior under dysbiotic circumstances and upregulation of host inflammatory responses. Among them, MMP-2 and MMP-9,[9] collectively called gelatinases, are morphologically similar to other members of the family except for distinct collagen binding sites, comprised fibronectin tandem repeats at the N-terminal region of the catalytic domain that favor gelatin binding.

The gelatinases demonstrate significance due to their ability to degrade type IV and type V collagen in the basement membrane, making them noteworthy in chronic inflammatory conditions such as periodontitis, arthritis, and cancer invasion. In periodontitis, MMP-2 and MMP-9 contribute substantially to the breakdown of gingival connective tissue. MMP-9[10] is found in early-stage inflammation in the gingival crevicular fluid and is produced by various inflammatory and noninflammatory cells, including fibroblasts, vascular smooth muscle cells, epithelial cells, osteoblasts, and keratinocytes. Meanwhile, platelets, endothelial cells, chondrocytes, osteoblasts, and other inflammatory cells principally secrete MMP-2. Activation of these latent enzymes occurs in the extracellular space or near the cell membrane.

Wang et al., in 2019, compared the chemically modified curcumin CMC 2.24 with natural curcumin and determined their potency as inhibitors of MMPs in cell culture influenced by lipopolysaccharides and in vivo rat models with periodontitis. They demonstrated that CMC 2.24 exhibited superior inhibition of MMP-9 secretion, with its natural counterpart being almost ineffective.[1,11] In addition, CMC 2.24 in vivo administration proved highly efficacious in suppressing the MMP-9 levels in plasma and the gingival tissue. It further indicated a decrease in systemic inflammatory status. Its effects were associated with decreased activation of nuclear factor kappa-light-chain-enhancer of activated B cells, which regulates gene products associated with inflammatory diseases.

Porphyromonas gingivalis can activate MMP-2 and MMP-9,[8] leading to tissue destruction and disease progression in periodontitis-affected patients. Treatment of aggressive periodontitis has shown a decrease in MMP-1,-8,-9,-12, and -13 in crevicular fluid, contributing to a destructive microenvironment. Reports have suggested that MMP-1,-8, and-13 are associated with enhanced alveolar bone resorption. Chemically modified curcumins,[11,12] like chemically modified curcumin-2.24, show potency to inhibit MMP-mediated degradation.

Drugs and genes[13] associated with MMP-9 are essential for disease diagnosis, prognosis, targeted therapies, personalized medicine, drug repurposing, and mechanistic studies. MMP-9 is involved in pathological processes such as chronic inflammation and cancer metastasis, which render it a significant biomarker in periodontal diagnostics. It also allows for the development of targeted drug therapies to manage periodontal diseases. Predicting drugs and genes associated with MMP-9 can help tailor treatments based on individual genetic profiles, optimizing treatment outcomes and minimizing side effects.[14] Predicting drug-gene associations is crucial for drug repurposing, personalized medicine, drug safety, mechanism of action elucidation, and drug discovery. It helps identify new targets for existing drugs, reduces time and cost for developing new therapeutics, and aids in personalized medicine by determining patient responses based on genetic profiles. Understanding interactions between drugs and genes is essential for assessing drug safety, guiding regulatory decisions, and improving patient safety. Predicting genes using deep learning algorithms associated with drugs and disease can lead to discovering new drug targets and more effective therapies. In one of our previous studies, we predicted drug-gene associations with periodontal pain using graph neural networks,[15] where SLC6A3, SLC6A2, FGF1, GRK2, and PLA2G2A were identified as central hub genes in P2X receptor-mediated drug-gene interactions. The GNN model achieved a 65% accuracy rate but demonstrated suboptimal predictive power for gene-drug interactions associated with oral pain.

Graph attention network (GAT) is a powerful tool for predicting drug-gene associations due to its ability to capture complex network structures and attention mechanisms. It can learn the relevance of each gene in interactions, focusing on the most informative nodes. GAT can handle heterogeneous information, enhancing prediction performance. It can handle large-scale networks and outperform traditional machine learning algorithms. This approach can aid in drug discovery, repurposing, and personalized medicine applications, providing accurate and reliable predictions. No studies have predicted drug-gene associations with MMP-9 for host modulation.

The study utilizes a GAT to predict drug-gene associations for MMP-9 in host modulation. MMP-9 protein is involved in disease processes like tissue remodeling and inflammation, allowing for the development of targeted therapies for cancer and inflammatory disorders. The research utilizes a GAT to accurately predict drug-gene associations, thereby contributing to precision medicine by identifying potential therapeutic targets for MMP-9 modulation in host modulation. Thus, our study aims to use GATs to predict drug-gene associations for MMP-9 in host modulation.

MATERIALS AND METHODS

Data preparation

Data on drugs and genes associated with MMP-9 were retrieved using the Probes and Drugs portal (www.probes-drugs.org).[16] Drug-gene interactions, totaling 1898, were collected from this site using the search icon. The data underwent cleaning to address missing values and to eliminate unnecessary columns. In preparation for graph data representation, the drug name and gene name were each represented by a node, while the target type was depicted as an edge. The edge weights reflect biochemical activity related to the nodes, whereas the node features provide additional information pertinent to the nodes for GAT training.

A total of 1898 drug–gene interactions related to MMP-9 were retrieved from the Probes and Drugs database for further analysis. The data were cleaned by removing any missing values and unnecessary columns. For the preparation of the graph data, drug names were represented as nodes, gene names were represented as nodes, and the target type was represented as edges. The edge weights represented the biochemical activity related to the nodes. In contrast, the node features provided additional details for training the GAT. Data were subjected to preprocessing steps for cleaning, removing missing values, and normalization.

Cytoscape

Using Cytoscape,[17] we created a network graph for drug-gene associations by importing data using a network file. We created a visualization and customized the layout with circular styles and labels, explored the network, analysed it using algorithms, and saved and exported the graph.

Cytohubba

Using Cytohubba, a drug-gene interaction network was imported, and the Maximum Clique Centrality (MCC) algorithm was applied. The MCC algorithm is based on clique enumeration and might not be scalable for large networks. Other methods, like degree, betweenness, or clustering coefficient, are recommended. Cytohubba generates a list of hub genes, which can be customized with parameters. Results were visualized and saved for further analysis.

Model architecture

Using Google Colab in a Python environment, a GAT[18] was applied. We used the GAT model [Figure 1], which consists of three layers, each designed to capture and aggregate information from neighboring nodes in the graph. The first layer is a GAT convolutional layer that takes the input node features and applies multiple attention heads to learn different aspects of the node’s neighbourhood. This layer outputs a set of hidden features for each node, which are then passed through an exponential linear unit (ELU) activation function to introduce nonlinearity. The second layer is another GAT convolutional layer that further processes the hidden features from the first layer, again using multiple attention heads to capture more complex patterns and relationships in the graph. This layer’s output is also passed through a ReLU activation function. The third and final layer is a GAT convolutional layer that reduces the hidden features to a single output feature per node, which is then passed through a SoftMax activation function to produce a probability score for binary classification.

Figure 1.

Graph attention network (GAT) architecture for predicting MMP-9 drug-gene associations. Schematic showing input features, attention mechanism, and multi-head attention layers in the GAT model. GAT – Graph attention network

Throughout the network, dropout is applied to prevent overfitting, and the model is trained using binary cross-entropy loss and the Adam optimizer. This architecture allows the model to effectively learn and aggregate information from the graph structure, making it well suited for tasks like node classification. The attention mechanism calculates attention coefficients for each connected node pair, which are normalized using the SoftMax function. The normalized attention coefficients are then aggregated to compute a weighted sum of neighbouring node features. GATs use multi-head attention to stabilize learning and enhance model expressiveness. The output layer generates final node embeddings or graph-level representations using a SoftMax function for class probabilities [Figure 1]. Hyperparameters used in this model are detailed in Table 1.

Table 1.

Training hyperparameters for the graph attention network model

Component	Setting/Value
Framework	Python (Google Colab)
Graph neural network type	GAT
Number of GAT layers	3
Layer 1: Attention heads	8
Layer 1: Activation	ELU
Layer 2: Attention heads	8
Layer 2: Activation	ReLU
Layer 3: Attention heads	1 (final aggregation)
Output layer activation	SoftMax
Loss function	Binary cross-entropy
Optimizer	Adam
Learning rate	0.001
Dropout rate	0.6
Weight initialization	Xavier uniform
Epochs	200
Batch size	Full batch (entire graph)
Evaluation metrics	Accuracy, F1 score, average precision
Training curve monitoring	Loss monitored over epochs

The table lists the configuration details of the GAT model including architecture, training parameters, and evaluation metrics. ELU – Exponential Linear Unit; ReLU – Rectified Linear Unit; SoftMax – Activation function converting outputs to probabilities; Binary Cross-Entropy – Loss function for binary classification; Adam – Adaptive Moment Estimation optimizer; Xavier Uniform – Weight initialization strategy; GAT – Graph attention network

RESULTS

The network graph has 742 nodes, 1897 edges, and an average number of neighbors of 5.049. The network diameter is 4, the longest shortest path is 4, and the network radius is 2. The characteristic path length is 3.303, indicating a relatively well-connected network. The clustering coefficient is 0.000, indicating low local connectivity. The network density is 0.007, indicating sparseness. Network heterogeneity is 2.801, indicating greater variation in the number of neighbors across nodes. Network centralization is 0.246, indicating a concentration of connections around a few central nodes. Two connected components indicate two distinct groups of nodes that are more connected internally. The analysis time is 0.366 s. The network graph is relatively well-connected, with a moderate average number of neighbors and characteristic path length. The top ten hubs identified with drug-gene associations with MMP-9 are Quercetin, Luteolin, Econazole, Zinc Chloride, Curcumin, MMP-9, MMP2, MMP1, MMP13, and MMP3 [Figure 2]. GATs model performance metrics [Table 1] include: Accuracy: 0.7955, F1 Score: 0.8861, and Average Precision: 0.9854.

Figure 2.

Drug-gene association network for matrix metalloproteinase-9. Visualization of known drug-gene associations. Nodes represent drugs or genes; edges represent validated associations. Constructed using cytoscape

The loss curve shows the model’s training progress. The initial high loss decreases rapidly, stabilizing and fluctuating around 0.47-0.48, indicating minor adjustments during training but converging to a stable state.

Loss values indicate a model’s training progress, which reflects its learning and parameter adjustments. The model’s stability, around 0.47–0.48, signifies a stable state. Performance metrics like accuracy, F1 score, and average precision are used to evaluate the model’s accuracy. The F1 score (0.8861) balances precision and recall, making it useful for dealing with imbalanced datasets.

Average precision (0.9854) measures the trade-off between precision and recall, particularly useful when the positive class is rare or when capturing the majority of positive instances. The model predicts all instances as positive, as evident from the high number of true positives (1509) and false positives (388) [Figure 3]. This bias toward the positive class suggests that the model may be overly sensitive or that there is a class imbalance issue in the dataset. The loss curve shows the model’s training progress, with rapid decreases in initial epochs indicating quick learning, stabilization indicating a suboptimal point, and normal fluctuations indicating parameter fine-tuning. The precision-recall curve [Figure 4] shows the trade-off between precision and recall for different classification thresholds. High precision indicates effective positive instance identification, but a sharp drop at high recall suggests a bias towards positive classes.

Figure 3.

Precision–recall curve of the graph attention network model. Evaluation of the Graph Attention Network model’s ability to predict drug-gene associations relevant to matrix metalloproteinases-9 modulation. Precision: The proportion of true positives among all predicted positives; Recall: The proportion of true positives among all actual positives. These metrics are crucial for assessing model effectiveness, particularly when false negatives and false positives have different implications

Figure 4.

Training loss curve of the Graph Attention Network model loss curve indicating model convergence during training. Lower loss values reflect better predictive accuracy

The model’s performance is hindered by a dataset imbalance, with 80% positive cases causing issues as it overfits the majority class. Despite these issues, the model learns useful features from the graph structure and shows stable training. The next steps should address this imbalance and improve the model’s discrimination ability. The model achieved an accuracy of 0.7955 (79.55%), indicating 80% correct classification. Its F1 score of 0.8861 indicates a good balance between precision and recall, and its average precision of 0.9854 is excellent.

The model’s performance improved during training, with the loss starting at 0.69 and dropping to about 0.47 in the first few epochs. After this initial improvement, the loss stabilized between 0.46 and 0.48, indicating that the model learned the most useful patterns early in training and made minor adjustments.

Precision measures the ratio of true positive predictions to the total number of positive predictions. In contrast, recall measures the ratio of true positive predictions to the total number of positive instances. The model’s precision is high at lower recall values, indicating confidence in positive predictions. However, it drops sharply at high recall values, indicating difficulty in maintaining precision and resulting in more false positives.

DISCUSSION

Matrix metalloproteinase (MMP) enzymes, including MMP-9, MMP2, MMP1, MMP13, and MMP3, are crucial in periodontal inflammation and host modulation. MMP-9 breaks down collagen and other extracellular matrix components, while MMP2 is involved in tissue repair. MMP1 breaks down interstitial collagens, contributing to collagen degradation and loss of connective tissue integrity. MMP13 is primarily expressed in osteoblasts and chondrocytes, contributing to bone and cartilage remodeling.[1,3,19] MMP3 degrades extracellular matrix components and is involved in tissue remodeling and inflammatory responses.

Top hub drug Quercetin[20,21] inhibits MMP-9 activity with an IC50 value of 22 μM, interacting with the S1′ subsite and affecting its inhibitory property through flavonoid R3′–OH and R4′–OH substitutions [Figure 2]. Quercetin treatment reduces MMP-2 and MMP-9 expressions in PC-3 cells, suggesting anticancer potential. It also treats diabetic retinopathy in rats by reducing MMP-9 and VEGF expression. Literature studies have reported their inhibitory roles in MMP-9 inhibition, proving their relevance in periodontal inflammation.

Luteolin, a plant compound, has shown potential as a treatment for human melanoma by inhibiting melanoma cell growth, increasing osteoblast cell counts, attenuating inflammation, and reducing inflammation in a Wistar rat experimental periodontitis model.[22] Curcumin inhibits AMPK and PKC pathways, reducing MMP-9, MMP-13, and EMMPRIN expression, providing insights into AMPK regulation and potential atherosclerosis treatment. Recent findings indicate that curcumin can effectively decrease the expression and activity of MMP-2 and MMP-9.

The present study aimed to elucidate the complex interplay between drugs and genes in the context of MMP-9 regulation. Employing a GAT (GAT),[23,24] the study was designed to identify potential therapeutic targets and understand the underlying mechanisms governing these interactions. The study findings reveal a network characterized by several central nodes, including quercetin, luteolin, econazole, zinc chloride, and curcumin, suggesting their pivotal roles in modulating MMP-9 activity. These compounds, positioned at the nexus of numerous drug-gene interactions, offer promising avenues for therapeutic intervention.

The GAT model’s performance in predicting drug-gene associations was encouraging, as evidenced by respectable accuracy, F1-score, and Average Precision [Table 1]. However, the inherent imbalance in the dataset, with a majority of positive instances, necessitates cautious interpretation of these metrics. Despite this limitation, identifying hub nodes significantly contributes to the field. Compounds like Quercetin and Luteolin, with established anti-inflammatory properties, have shown promise in preclinical studies, and their emergence as central network nodes strengthens their potential therapeutic utility.[25,26]

A recent study proposed a GAT framework[27,28] which predicts piRNA disease association and virus-drug association, outperforming collaborative filtering and attribute feature-based methods. This demonstrates the potential of graph neural networks in biomedical research, with an accuracy of 0.9038, similar to our study’s accuracy of 79% [Figure 3].

Techniques like oversampling or undersampling can create a balanced dataset to address dataset imbalance. Advanced approaches such as data augmentation or synthetic data generation may improve model performance. Fine-tuning hyperparameters such as learning rate, regularization methods, and network architecture can enhance discrimination ability. GATs, which rely on graph structure, may require significant computational resources for larger datasets. Generalizability is crucial for learning features specific to the current dataset.[29,30]

A key strength of this study is its innovative use of a GAT to explore how drugs interact with associated genes in regulating MMP-9. We identified important hub nodes and potential therapeutic targets by creating a detailed network of these interactions. Using a large dataset, we discovered new associations and developed testable hypotheses, accelerating the drug discovery process. Highlighting quercetin and luteolin as central nodes points to new treatment strategies and supports repurposing existing drugs. This blend of computational modeling and biological knowledge offers unique and valuable insights into diseases related to MMP-9.

While the current study’s findings offer valuable insights, it is important to acknowledge their limitations. By focusing solely on MMP-9, we have overlooked the complex network of drug-gene interactions involved in this biological process. In addition, the observational nature of our study prevents establishing definitive cause-and-effect relationships. Moreover, the availability of comprehensive drug-gene interaction data could influence the scope of our network analysis.[2,10,14,31]

To tackle the issue of class imbalance, we will implement strategies including the Synthetic Minority Oversampling Technique, adjusting class weights during model training, and reducing the size of the majority class. These strategies aim to enhance the model’s overall predictive accuracy while improving its ability to correctly identify true negatives. Furthermore, to better evaluate the model’s performance with imbalanced data, we will include ROC-AUC and recall metrics in forthcoming assessments. This approach will allow a more comprehensive evaluation of the model’s effectiveness in detecting all possible drug-gene associations.

Future research should explore a broader range of MMP proteins and incorporate a wider array of biological data to build upon these findings. Experimental validation of the predicted drug-gene interactions is crucial to solidify the conclusions of the present study. Advanced machine learning techniques and multi-omics data integration could also provide fresh perspectives. Translating these research findings into tangible patient benefits will require rigorous clinical trials.

CONCLUSION

This study provides a foundation for understanding the intricate relationship between drugs, genes, and MMP-9. The application of a GAT has proven to be a valuable tool in identifying potential drug targets and unravelling the complexities of these interactions. By addressing the identified limitations and expanding the scope of research, we can significantly advance our understanding of MMP-9 biology and develop novel therapeutic strategies for diseases characterized by MMP-9 dysregulation.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.